Hello, everyone—I’m staying involved here, too.
Innodata is releasing the first phase of its AI cyber training suite to enable AI coding agents to write—and fix—secure code.
NEW YORK, NY / ACCESS Newswire / August 4, 2026 / INNODATA INC. (Nasdaq:INOD) today launched the first phase of its AI Cyber Training Suite —twelve datasets and evaluation systems that train AI coders to write code that avoids known security vulnerabilities, introduces new ones, and fixes existing vulnerabilities in software systems already in operation. The suite now addresses a key hurdle to trust in AI-written code: the risk that an agent, while adding a feature or modernizing a legacy system, might inadvertently create a security vulnerability. The AI Cyber Training Suite works across the entire stack of coding agents—the models, harnesses, and tools—to ensure that the code they generate is not only functional but also secure. It is available today for model builders and enterprises.
The suite is based on thousands of real-world security vulnerabilities discovered over the past decade in Python, TypeScript, JavaScript, Rust, C, Go, and other languages, and covers a wide range of platforms and service integrations, including Linux, macOS, Android, Windows, AWS, GCP, and many others. Each vulnerability was manually dissected by Innodata’s cybersecurity engineers and reconstructed using a targeted attack in a sealed, offline copy of the vulnerable software. This approach makes the suite unusually rigorous: it measures an AI’s ability to fix a vulnerability by verifying that the original attack against the patched code is no longer successful, while simultaneously confirming that the software continues to function as intended. The result is data and metrics that both train coding agents to avoid vulnerabilities from the outset and test their ability to fix existing vulnerabilities.
"For model builders, these evaluations and datasets enable the coding agents—the models, harnesses, and tools—to ensure that the code they generate is not only effective but also secure,” said Yonatan Brander, who leads the cybersecurity data engineering team at Innodata. "And for companies, they offer something equally important: the confidence that when teams refactor, add features, or modernize legacy systems with AI agents, they won’t reintroduce the very vulnerabilities they had already avoided or patched."
For companies, this means they can use AI to modernize legacy systems while minimizing the risk of security vulnerabilities and leveraging AI to fix identified vulnerabilities.
Innodata tested the ability of ten leading AI models and agents to fix verified security vulnerabilities and found that the best Openweight AI models fixed at most 23% of these verified flaws—even when they were shown exactly where each one was located. However, after a single round of fine-tuning using a subset of the suite’s data, the model’s ability to fix vulnerabilities without being prompted more than doubled to 41.2%—rising from 18.4% to 41.2%.
“AI is now writing a growing share of the world’s code, and the agents writing it must ensure that this code is secure,” said Jack Abuhoff, CEO of Innodata. “Otherwise, companies cannot rely on AI-generated code without a professional security team reviewing everything that is produced—and that is a real barrier to AI adoption. The Cyber OTS Suite helps companies confidently deploy agent-generated code into production.”
